Datagedreven Voorspelling en Regeling van Residentiële PV-Batterijsystemen Voorbij Perfecte Voorkennis

The electrification of the residential sector is turning households into complex energy hubs that combine rooftop photovoltaics (PV), battery energy storage systems (BESSs), electric vehicles (EVs), heat pumps (HPs), and flexible demand. While these technologies offer opportunities to reduce costs and emissions, their real-world performance depends on operational control, tariff choice, and user behaviour. Most existing studies on PV-battery design rely on idealised assumptions—perfect foresight, deterministic inputs, or simplified control—which leads to large gaps between predicted and achievable performance. This dissertation addresses these limitations by developing a methodological framework that evaluates residential energy systems under realistic forecast uncertainty. The core objective is to quantify how operational realism—imperfect forecasts, finite-horizon model predictive control (MPC), and tariff complexity—affects the economic and environmental performance of residential energy systems. To this end, the dissertation develops a hybrid cyber-physical co-simulation framework that integrates high-resolution household simulations, a modular forecasting benchmark (EnFoBench), and three control approaches: rule-based control (RBC), finite-horizon MPC, and full-horizon mixed-integer linear programming (MILP). The framework is applied to Belgian households, both with and without electric vehicles (EVs), and evaluated under fixed, monthly varying, hourly dynamic tariffs with the inclusion of the recently introduced capacity tariffs. This unified setup enables consistent and fair comparison of operational strategies under realistic conditions. Forecasts are produced using statistical, machine-learning, and foundation models, and their influence is assessed through key performance indicators including energy cost, emissions, grid interaction, and return on investment (ROI). The first major finding concerns the comparative performance of operational control strategies. Across all households and tariff structures, full-horizon MILP establishes the theoretical upper bound on achievable performance. In contrast, realistic implementations of MPC deliver only a fraction of this potential: even with perfect forecasts they capture just 75 % to 95 % of the MILP optimum, and once forecast errors are introduced their performance deteriorates sharply. Under realistic conditions, MPC rarely outperforms simple RBC. These results demonstrate that the expected advantages of predictive optimisation largely disappear once finite horizons and imperfect forecasts are taken into account. The second major finding relates to the influence of forecasting. Forecast quality—both in aggregate accuracy and in the temporal structure of errors—emerges as a dominant determinant of operational cost. Load-forecast errors have particularly strong cost impacts under capacity-based tariffs, and their effects compound with PV-forecast errors. Crucially, more frequent re-forecasting and re-optimisation does not guarantee better outcomes: updating more often than once per day frequently amplifies short-term noise, destabilises optimisation routines, and worsens economic performance. The results show that forecast stability and coherence are as important as numerical accuracy when forecasts are embedded in closed-loop control. The third major finding centres on the role of electric vehicles. EV charging introduces high-power, stochastic demand peaks that are difficult to anticipate under forecast uncertainty, significantly complicating household operation. In households without EVs, batteries can achieve modest positive ROI under realistic control; however, the introduction of an EV reduces these returns substantially. Both RBC and realistic MPC often fail to mitigate EV-driven peaks, leading to outcomes that are close to break-even or, in some cases, costlier than having no battery at all. Only the idealised MILP benchmark maintains strong performance, highlighting the gap between theoretical potential and practical value once EV behaviour is included. Taken together, these findings show that operational realism—forecast uncertainty, controller limitations, and EV behaviour—is the primary determinant of achievable economic value in residential PV-battery systems. They explain why deterministic design studies consistently overestimate benefits, and they underscore the need to treat forecasting and control assumptions as integral components of techno-economic assessment.